Open-weight AI promises more control. What’s the catch?
Open-weight AI is expanding enterprise choice — and changing how leaders evaluate cost, control and risk.
Open-weight AI models are creating new possibilities for how organizations deploy, customize and manage artificial intelligence. Unlike proprietary models, which are typically accessed through a provider, open-weight models give enterprises the option to run AI within their own infrastructure.
For business leaders, that flexibility raises important questions about cost, sovereignty, security and long-term strategy. But the conversation shouldn’t be about whether open-weight or proprietary AI models are better. It should be about understanding which model can deliver the greatest value to an organization, including meeting its security requirements. Organizations may also turn to specialized models, whether open or closed, to achieve their goals.
Here, Giovanni Carraro, Senior Vice President, Global Strategic Alliances, discusses the trade-offs between open and closed models as more firms move from experimentation to enterprise-wide AI adoption, and how Kyndryl can help customers navigate their AI partner choices.
Why are open-weight AI models gaining attention among enterprise leaders?
Giovanni Carraro: The conversation about enterprise AI is evolving. Organizations are moving beyond experimentation and thinking more seriously about how AI fits into their operations, infrastructure and long-term business strategies.
Cost, control and customization are driving much of the interest, and some organizations are finding that open-weight models can be a part of their enterprise tools.
Open-weight models allow organizations to access, download and customize trained AI models to meet specific business needs, giving them greater control over where AI runs, how data is managed and the cost model under which it operates. For enterprises looking to scale AI adoption, that flexibility may become increasingly valuable as they balance growing demands with operational and financial considerations.
How should enterprises think about open-weight versus closed, proprietary AI models?
Carraro: This shouldn’t be an either-or decision. Both have advantages, and the right choice depends on an organization’s needs and how it plans to use AI.
There are differences in how you access and deploy these models. It’s helpful to think about “surfaces” and “engines.” Surfaces are where you access the tool; engines are the AI model the surface directs to.
Proprietary models are typically accessed through web, desktop, and mobile applications; through APIs, either directly from the developer or via managed cloud platforms; through command-line tools and IDE extensions; and increasingly through integrations embedded in enterprise software. In these instances, the AI provider manages much of the underlying infrastructure, which can simplify deployment and ongoing maintenance for the end user. Proprietary models can still offer choices of hosting location and capacity, and in some cases can offer fine-tuning, within the model provider’s available options.
Open-weight models can offer greater flexibility over where and how AI operates. Organizations can host them in private cloud or on-premises environments, or even on individual devices, and can customize certain models for specialized tasks. Some open providers offer models through cloud services, which means they would manage the infrastructure and maintenance.
It’s also important to distinguish open-weight from open-source AI. While open-weight models provide access to trained model weights, they don’t necessarily include the underlying source code or training datasets that may be available with open-source AI. That distinction matters because the ability to run or customize a model doesn’t necessarily mean an organization has full visibility into how it was developed, whether it is free of unwanted behaviors, or what biases it may carry. That requires rigorous testing.
Ultimately, the question isn’t which model is universally better. It’s which model is best suited to an organization’s specific needs and business objectives.
Can open-weight AI models help enterprises manage costs?
Carraro: In short, yes. Open-weight models can offer more cost-effective alternatives to proprietary models, particularly for organizations that rely heavily on AI or use it for specific business functions.
With proprietary models, businesses generally pay for access and consumption. Open-weight models give organizations more flexibility in how they deploy AI, potentially reducing ongoing costs associated with accessing commercial models.
However, downloading a model for free doesn’t eliminate the cost of running it. It shifts the cost elsewhere, from per-token to infrastructure and maintenance. Enterprises still need computing resources, proper infrastructure, security controls and the expertise to manage and maintain them.
Whether open or closed, a smaller, more tailored AI model may be more cost-effective than a larger general-purpose model for an organization that only needs it to perform a specialized or less complex task. The focus should be on overall cost and business value, versus whether the model is available to download freely.
The goal isn’t to commit to a particular model: it’s to make informed decisions that support business objectives, deliver value and provide the flexibility to adapt.
What should enterprises prioritize as they develop their AI model strategies?
Carraro: Flexibility should be a priority from the beginning.
Organizations are increasingly exploring approaches that combine different models for different business needs. An enterprise might use a proprietary model for one application and a smaller, specialized open-weight model for another.
That makes interoperability important. Businesses should consider how easily they can introduce, replace or manage models as their needs and the technology evolve.
This is also where Kyndryl Consult and our technology experts can help organizations evaluate their priorities, understand the trade-offs between different technologies and consider how AI choices align with their existing IT environments.
At Kyndryl, our focus is on supporting the best option for specific business needs. The goal isn’t to commit to a particular model: it’s to make informed decisions that support business objectives, deliver value and provide the flexibility to adapt.
What role does data sovereignty play in the growing interest in open-weight AI?
Carraro: Digital sovereignty is becoming an increasingly important consideration, especially for organizations operating across different regulatory environments.
Open-weight models can give enterprises greater control over where models run and where data is processed. For example, an organization may choose to deploy a model within its own IT infrastructure or in a sovereign cloud environment.
That flexibility can be valuable when data residency, privacy or operational control are priorities.
But hosting a model locally doesn’t automatically resolve every sovereignty or compliance concern. Organizations still need to evaluate the model’s security, its configuration and its interactions with other systems.
Does greater control also mean greater security responsibility? How do you know if an open model is safe?
Carraro: Greater flexibility often comes with greater operational responsibility.
Proprietary AI models generally place more responsibility for safeguards and security controls with the provider. With open-weight models, organizations have greater control, but also more responsibility for monitoring and managing the models. That said, greater control can allow organizations to conduct their own testing and carry out certain cybersecurity activities that proprietary models may restrict.
Regardless of the model they choose, enterprises need to consider whether models come from trusted sources, who has access to them, how data is protected and how the models are monitored. Open models should be evaluated for safety regardless of their country of origin or developer. The important point is that security and governance must be built into the broader AI operating environment.